def main():
parser = argparse.ArgumentParser(
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
# Path options.
parser.add_argument("--pretrained_model_path",
default=None,
type=str,
help="Path of the pretrained model.")
parser.add_argument("--dataset_path_list",
default=[],
nargs='+',
type=str,
help="Dataset path list.")
parser.add_argument("--output_model_path",
default="models/multitask_classifier_model.bin",
type=str,
help="Path of the output model.")
parser.add_argument("--config_path",
default="models/bert/base_config.json",
type=str,
help="Path of the config file.")
# Model options.
model_opts(parser)
parser.add_argument("--pooling",
choices=["mean", "max", "first", "last"],
default="first",
help="Pooling type.")
# Tokenizer options.
tokenizer_opts(parser)
# Optimizer options.
optimization_opts(parser)
# Training options.
training_opts(parser)
adv_opts(parser)
args = parser.parse_args()
args.soft_targets = False
# Load the hyperparameters from the config file.
args = load_hyperparam(args)
set_seed(args.seed)
# Count the number of labels.
args.labels_num_list = [
count_labels_num(os.path.join(path, "train.tsv"))
for path in args.dataset_path_list
]
args.datasets_num = len(args.dataset_path_list)
# Build tokenizer.
args.tokenizer = str2tokenizer[args.tokenizer](args)
# Build multi-task classification model.
model = MultitaskClassifier(args)
# Load or initialize parameters.
load_or_initialize_parameters(args, model)
# Get logger.
args.logger = init_logger(args)
args.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = model.to(args.device)
args.model = model
if args.use_adv:
args.adv_method = str2adv[args.adv_type](model)
# Training phase.
dataset_list = [
read_dataset(args, os.path.join(path, "train.tsv"))
for path in args.dataset_path_list
]
packed_dataset_list = [
pack_dataset(dataset, i, args.batch_size)
for i, dataset in enumerate(dataset_list)
]
packed_dataset_all = []
for packed_dataset in packed_dataset_list:
packed_dataset_all += packed_dataset
instances_num = sum([len(dataset) for dataset in dataset_list])
batch_size = args.batch_size
args.train_steps = int(instances_num * args.epochs_num / batch_size) + 1
args.logger.info("Batch size: {}".format(batch_size))
args.logger.info(
"The number of training instances: {}".format(instances_num))
optimizer, scheduler = build_optimizer(args, model)
if args.fp16:
try:
from apex import amp
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use fp16 training."
)
model, optimizer = amp.initialize(model,
optimizer,
opt_level=args.fp16_opt_level)
args.amp = amp
if torch.cuda.device_count() > 1:
args.logger.info("{} GPUs are available. Let's use them.".format(
torch.cuda.device_count()))
model = torch.nn.DataParallel(model)
total_loss, result, best_result = 0.0, 0.0, 0.0
args.logger.info("Start training.")
for epoch in range(1, args.epochs_num + 1):
random.shuffle(packed_dataset_all)
model.train()
for i, (dataset_id, src_batch, tgt_batch,
seg_batch) in enumerate(packed_dataset_all):
if hasattr(model, "module"):
model.module.change_dataset(dataset_id)
else:
model.change_dataset(dataset_id)
loss = train_model(args, model, optimizer, scheduler, src_batch,
tgt_batch, seg_batch, None)
total_loss += loss.item()
if (i + 1) % args.report_steps == 0:
args.logger.info(
"Epoch id: {}, Training steps: {}, Avg loss: {:.3f}".
format(epoch, i + 1, total_loss / args.report_steps))
total_loss = 0.0
for dataset_id, path in enumerate(args.dataset_path_list):
args.labels_num = args.labels_num_list[dataset_id]
if hasattr(model, "module"):
model.module.change_dataset(dataset_id)
else:
model.change_dataset(dataset_id)
result = evaluate(
args, read_dataset(args, os.path.join(path, "dev.tsv")))
save_model(model, args.output_model_path)
def main():
parser = argparse.ArgumentParser(
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
finetune_opts(parser)
parser.add_argument(
"--max_choices_num",
default=4,
type=int,
help=
"The maximum number of cadicate answer, shorter than this will be padded."
)
tokenizer_opts(parser)
adv_opts(parser)
args = parser.parse_args()
args.labels_num = args.max_choices_num
# Load the hyperparameters from the config file.
args = load_hyperparam(args)
set_seed(args.seed)
# Build tokenizer.
args.tokenizer = str2tokenizer[args.tokenizer](args)
# Build multiple choice model.
model = MultipleChoice(args)
# Load or initialize parameters.
load_or_initialize_parameters(args, model)
# Get logger.
args.logger = init_logger(args)
args.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = model.to(args.device)
# Training phase.
trainset = read_dataset(args, args.train_path)
instances_num = len(trainset)
batch_size = args.batch_size
args.train_steps = int(instances_num * args.epochs_num / batch_size) + 1
args.logger.info("Batch size: {}".format(batch_size))
args.logger.info(
"The number of training instances: {}".format(instances_num))
optimizer, scheduler = build_optimizer(args, model)
if args.fp16:
try:
from apex import amp
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use fp16 training."
)
model, optimizer = amp.initialize(model,
optimizer,
opt_level=args.fp16_opt_level)
args.amp = amp
if torch.cuda.device_count() > 1:
args.logger.info("{} GPUs are available. Let's use them.".format(
torch.cuda.device_count()))
model = torch.nn.DataParallel(model)
args.model = model
if args.use_adv:
args.adv_method = str2adv[args.adv_type](model)
total_loss, result, best_result = 0.0, 0.0, 0.0
args.logger.info("Start training.")
for epoch in range(1, args.epochs_num + 1):
random.shuffle(trainset)
src = torch.LongTensor([example[0] for example in trainset])
tgt = torch.LongTensor([example[1] for example in trainset])
seg = torch.LongTensor([example[2] for example in trainset])
model.train()
for i, (src_batch, tgt_batch, seg_batch,
_) in enumerate(batch_loader(batch_size, src, tgt, seg)):
loss = train_model(args, model, optimizer, scheduler, src_batch,
tgt_batch, seg_batch)
total_loss += loss.item()
if (i + 1) % args.report_steps == 0:
args.logger.info(
"Epoch id: {}, Training steps: {}, Avg loss: {:.3f}".
format(epoch, i + 1, total_loss / args.report_steps))
total_loss = 0.0
result = evaluate(args, read_dataset(args, args.dev_path))
if result[0] > best_result:
best_result = result[0]
save_model(model, args.output_model_path)
# Evaluation phase.
if args.test_path is not None:
args.logger.info("Test set evaluation.")
if torch.cuda.device_count() > 1:
args.model.module.load_state_dict(
torch.load(args.output_model_path))
else:
args.model.load_state_dict(torch.load(args.output_model_path))
evaluate(args, read_dataset(args, args.test_path))
def main():
parser = argparse.ArgumentParser(
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
finetune_opts(parser)
parser.add_argument("--pooling",
choices=["mean", "max", "first", "last"],
default="first",
help="Pooling type.")
tokenizer_opts(parser)
parser.add_argument("--soft_targets",
action='store_true',
help="Train model with logits.")
parser.add_argument("--soft_alpha",
type=float,
default=0.5,
help="Weight of the soft targets loss.")
args = parser.parse_args()
# Load the hyperparameters from the config file.
args = load_hyperparam(args)
set_seed(args.seed)
# Count the number of labels.
args.labels_num = count_labels_num(args.train_path)
# Build tokenizer.
args.tokenizer = str2tokenizer[args.tokenizer](args)
# Build classification model.
model = Classifier(args)
# Load or initialize parameters.
load_or_initialize_parameters(args, model)
args.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = model.to(args.device)
# Training phase.
trainset = read_dataset(args, args.train_path)
random.shuffle(trainset)
instances_num = len(trainset)
batch_size = args.batch_size
src = torch.LongTensor([example[0] for example in trainset])
tgt = torch.LongTensor([example[1] for example in trainset])
seg = torch.LongTensor([example[2] for example in trainset])
args.train_steps = int(instances_num * args.epochs_num / batch_size) + 1
print("Batch size: ", batch_size)
print("The number of training instances:", instances_num)
optimizer, scheduler = build_optimizer(args, model)
if args.fp16:
try:
from apex import amp
except ImportError:
raise ImportError(
"Please install apex from https://www.github.com/nvidia/apex to use fp16 training."
)
model, optimizer = amp.initialize(model,
optimizer,
opt_level=args.fp16_opt_level)
args.amp = amp
if torch.cuda.device_count() > 1:
print("{} GPUs are available. Let's use them.".format(
torch.cuda.device_count()))
model = torch.nn.DataParallel(model)
args.model = model
total_loss, result, best_result = 0.0, 0.0, 0.0
print("Start training.")
for epoch in range(1, args.epochs_num + 1):
model.train()
for i, (src_batch, tgt_batch, seg_batch,
_) in enumerate(batch_loader(batch_size, src, tgt, seg)):
loss = train_model(args, model, optimizer, scheduler, src_batch,
tgt_batch, seg_batch)
total_loss += loss.item()
if (i + 1) % args.report_steps == 0:
print("Epoch id: {}, Training steps: {}, Avg loss: {:.3f}".
format(epoch, i + 1, total_loss / args.report_steps))
total_loss = 0.0
result = evaluate(args, read_dataset(args, args.dev_path))
if result > best_result:
best_result = result
save_model(model, args.output_model_path)
# Evaluation phase.
if args.test_path is not None:
print("Test set evaluation.")
if torch.cuda.device_count() > 1:
args.model.module.load_state_dict(
torch.load(args.output_model_path))
else:
args.model.load_state_dict(torch.load(args.output_model_path))
evaluate(args, read_dataset(args, args.test_path))